Dialogue Generation on full-hh-rlhf (test)
79.3Win Rate (Beaver-7b-v3.0-reward)ReMax+XRLHF
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ReMax+XRLHFBase Model=pythia-2.8B2025.12 | 79.3 | 80.1 | |
| ReMax+XRLHFBase Model=opt-1.3B2025.12 | 77.2 | 76.1 | |
| PPO+XRLHFBase Model=opt-1.3B2025.12 | 76.4 | 78.5 | |
| PPO+XRLHFBase Model=pythia-2.8B2025.12 | 75.8 | 76.5 | |
| ReMaxBase Model=pythia-2.8B2025.12 | 71.4 | 66.8 | |
| ReMaxBase Model=opt-1.3B2025.12 | 70.6 | 66.9 | |
| PPOBase Model=pythia-2.8B2025.12 | 69.8 | 67.5 | |
| PPOBase Model=opt-1.3B2025.12 | 68.3 | 68 |